Krea Releases 12-Billion Parameter Text-to-Image Model with 8-Step Generation

Krea.ai released Krea 2 Turbo on June 22, 2026, a 12-billion parameter diffusion transformer model for text-to-image generation. The open-weight model generates images in 8 inference steps and supports resolutions up to 2048x2048 pixels.

Model Specifications

Krea 2 uses a diffusion transformer architecture with 12 billion parameters. The model family includes two checkpoints: Krea 2 Raw (base model) and Krea 2 Turbo (post-trained with fine-tuning and distillation). According to Krea, the Turbo variant requires only 8 inference steps with a guidance scale of 0.0.

The model runs on the Hugging Face diffusers library and SGLang inference framework. Default inference parameters include mu=1.15 for the official codebase and support for resolutions including 1024x1024 and 2048x2048 pixels.

Training and Data

Krea.ai states the model was trained on publicly available data, licensed third-party data, and proprietary synthetic data. The company claims data was filtered to remove harmful content and low-quality duplicates before training. Specific dataset size and composition details were not disclosed.

Safety Implementation

The model operates under the Krea 2 Community License, which requires deployers to implement content filtering or review processes. Krea reports conducting internal and external safety evaluations including adversarial testing for sexually explicit content, non-consensual intimate imagery, and child safety risks before release.

For Krea's hosted products, the company deploys input and output classifiers using proprietary and third-party detection tools. The open-weight release does not include these safeguards—deployers must implement their own filtering systems per license requirements.

Availability

The model is available on Hugging Face as an open-weight release. The repository includes turbo.safetensors weights and integration code for diffusers and SGLang. Krea also offers hosted product integrations, though pricing was not disclosed.

What This Means

Krea 2 Turbo enters a competitive text-to-image market with a relatively large 12-billion parameter model that claims 8-step generation—faster than many diffusion models that require 20-50 steps. The open-weight release with license-mandated safety requirements represents a middle path between fully open and closed models, though enforcement of deployer compliance remains unclear. The model's actual image quality and prompt-following capabilities compared to established models like Stable Diffusion 3 and DALL-E 3 will depend on independent benchmarking.